context-degradation

Context degradation occurs when agent performance declines as context grows, manifesting as lost-in-middle effects, poisoning, or distraction. It reduces attention to relevant information and compounds in long conversations, debugging outputs, and production systems handling large contexts, requiring detection and mitigation strategies.

Use the degradation_detector.py script to analyze your agent's context by calling analyze_agent_context() with your context string. The tool measures attention distribution, quantifies degradation metrics like lost-in-middle and poisoning patterns, and returns a health score with recommended recovery steps.

Context degradation mitigation includes compaction to reduce noise, masking to isolate critical information, and isolation to separate poisoned or distracting content. The Skill provides actionable guidance on which strategy fits your degradation pattern and conversation state.

Monitor context health continuously in production agents handling long-running conversations or large contexts to catch degradation early. Degradation typically worsens over time, so regular benchmarking preserves performance before reliability issues compound.

Yes, the Skill applies to any agent system working with long contexts. It requires only numpy and your agent's current context string, making it compatible with debugging workflows, benchmarking pipelines, and existing large-context architectures without framework-specific modifications.